Early diagnosis of ATTR cardiomyopathy by AI-assisted real-time screening of incidental cardiac uptake on bone scintigraphy

Background Transthyretin (ATTR) cardiomyopathy is an underdiagnosed but treatable condition, in which early detection is critical for patient outcomes. ATTR can be incidentally detected on oncologic bone scintigraphy as cardiac uptake. In particular, subtle cardiac uptake might be overlooked and requires confirmation by single-photon emission tomography (SPECT). We developed and implemented an artificial intelligence (AI)-based workflow to enable real-time detection of cardiac uptake and to study the possibility of integrating it into routine clinical practice. Methods The workflow consists of 1) acquisition of scintigraphy, 2) automatic exporting of images to the analysis server, 3) preprocessing and classification of images with a convolutional neural network (CNN), 4) an automated notification system. The feasibility and accuracy of the AI workflow were evaluated in retrospective (n = 11 293) and prospective (n = 393) bone scintigraphy cohorts. A nuclear medicine (NM) physician validated all patients using the Perugini grade (grades 0–1 negative and 2–3 positive). Results In prospective real-time use (n = 393), the workflow resulted in 11 false positives and 11 true positives. No false negative findings were observed, achieving an area under the curve (AUC) of 0.99. In the retrospective cohort (n = 11293), the AI correctly classified 31 true positives and 10949 negative images. However, the AI classification yielded 308 false positives and 5 false negatives (AUC of 0.98). Conclusion The automated CNN-based detection algorithm was feasible in a real-time clinical workflow for the opportunistic screening of ATTR cardiomyopathy. The workflow helps to ensure a successful diagnostic loop for incidental cardiac uptake through automated detection, immediate radiographer notification, and the option for further SPECT imaging. However, the positive result always requires expert validation. We demonstrate how a highly sensitive, scalable framework for AI could be integrated into routine imaging workflows for the timely diagnosis of ATTR. Shared codes support the reproducibility of results and further development.